Download Constitutive Rules in Game Theory: Two Accounts, One Rejection

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Game mechanics wikipedia , lookup

Deathmatch wikipedia , lookup

The Evolution of Cooperation wikipedia , lookup

Rules of chess wikipedia , lookup

Artificial intelligence in video games wikipedia , lookup

Evolutionary game theory wikipedia , lookup

Chicken (game) wikipedia , lookup

Transcript
Constitutive Rules in Game Theory:
Two Accounts, One Rejection
Cyril Hédoin
REGARDS - University of Reims Champagne-Ardenne
First draft: 24/03/2013
Abstract: Game theory and rules are deeply intertwined for at least two reasons: first, in
many cases rules are necessary to break the indeterminacy that surrounds most of the games;
second, in the past thirty years game theory has been increasingly used as a major tool to
build a theory of social rules. Interestingly, though the concept of rules is now part of most
game theorists’ tool box, none of them has explicitly entertained the important distinction
between regulative rules and constitutive rules. This distinction, which finds its roots in
Ludwig Wittgenstein’s profound discussion of “language games”, is at the core of modern
philosophy. This article asks whether game theory can account for constitutive rules. I
distinguish between three game-theoretic accounts of rules: the rules-as-behavioralregularities account, the rules-as-normative-expectations account and the rules-as-correlateddevices account. I show that only the latter two are able make sense of constitutive rules by
giving up the individualistic understanding of game theory.
Keywords: constitutive rules, psychological game theory, epistemic game theory,
methodological individualism
Word count: 12649

Paper prepared for the XI. INEM Conference at Rotterdam, 13-15 June 2013.
Email: [email protected]
1
Constitutive Rules in Game Theory:
Two Accounts, One Rejection
1. Introduction
At the most general level, game theory can be defined as the theory of rational decision in
strategic interactions, i.e. decision problems where the consequences of all agents’ decisions
are a function of others’ decisions. Usually, at least in its classical (non-evolutionary) version,
game theory endows agents with two forms of rationality (Bicchieri 1993): practical
rationality assumes that agents choose whatever is best for them according to their preferences
and their beliefs; epistemic rationality guarantees that the agents’ beliefs are consistent
between them and with the available information. The systematic study of the behavior of
rational agents in both the practical and epistemic meanings has revealed unanticipated
difficulties preventing game theorists to establish a fully predictive and explanatory theory of
rational agents in strategic interactions. Undoubtedly, one of the most serious difficulty has
been (and still is) the indeterminacy problem ((Bicchieri 2009); (Hargreaves-Heap &
Varoufakis 2004)): in several and significant kinds of games, the result of the interaction
between rational agents cannot be predicted because of the lack of solution concept pointing
to a unique outcome. Assuming that rational agents know the theory of games as well as game
theorists, this means that agents themselves should be unable to settle on a specific strategy.
Obviously, facts are not consistent with this stark conclusion.
Indeterminacy is one of the major reasons why many consider game theory to be a failure
both as a descriptive (positive) and as a prescriptive (normative) theory. To overcome
indeterminacy is thus essential. One recent strategy has been to introduce in the analysis
social objects such as “institutions”, “social norms”, “conventions” or “rules”. Social objects
break indeterminacy because they select one equilibrium1 among the many that may coexist in
a game. For instance, the fact that everyone in France is driving at the right side of the road is
explained by the fact that it is settled by a legal rule that everyone knows and which everyone
expects others to know and to follow.2 The obvious drawback of this strategy is that it seems
plainly ad hoc: one has solved a foundational difficulty of game theory by simply introducing
a new concept with no game-theoretical foundation. This may explain why at least for the last
thirty years game theory has became also a major tool in the construction of a theory of social
rules. One way to escape the critique of “ad hoc-ness” has been simply to make institutions
and social rules endogenous by explaining their emergence and their maintenance as the
product of rational choices in strategic interactions: social rules (partially) determine agents’
choices, and in return these choices sustain the rules in a self-fulfilling process.
The amount of works in the social sciences using game theory to study institutions and social
rules is now huge and keeps growing at a fast pace. Interestingly, though the concept of rules
1
Unless it is specified differently, in this paper I always use the term “equilibrium” as a synonym for the Nash
equilibrium solution concept. An outcome is a Nash equilibrium if every player is playing his best response
given the strategies of all the other players.
2
In some game-theoretic accounts of institutions, it is required to the social object to be common knowledge.
See for example David Lewis’ theory of conventions (Lewis 1969).
2
is now part of most game theorists’ tool box, none of them has explicitly entertained the
important distinction between regulative rules and constitutive rules. This distinction, which
finds its roots in Ludwig Wittgenstein’s profound discussion of “language games”, is at the
core of modern philosophy. More recently, it has been particularly advanced by philosophers
such as John Rawls or John Searle to account respectively for the nature of moral doctrines
and of the social world. While regulative rules merely regulate existing practices, constitutive
rules define social practices which could not exist without them. Assuming that game theory
has no difficulty to account for the former, one can wonder whether constitutive rules may be
analyzed within a game-theoretic framework. This is the question that will interest me in the
rest of paper. I single out two game-theoretic accounts relevant to produce a relevant study of
constitutive rules: an account based on psychological game theory and an account grounded
on the solution concept of correlated equilibrium. As it will be clear, this issue has
foundational implications for game theory. Indeed, I claim that to account for constitutive
rules, any individualistic understanding of game theory must be given up. This claim
reinforces Herbert Gintis’ (2009) thesis that game theory is incompatible with methodological
individualism.
The rest of the paper is divided into six sections. The next section defines more precisely the
concept of constitutive rules and contrasts it with regulative rules. It briefly indicates why an
understanding of constitutive rules is important for economists and game-theorists. The third
section shows that the standard account of social rules in game theory is unable to account for
constitutive rules. The fourth and fifth sections present two relatively new approaches that I
claim to be relevant given the issue at stake: Cristina Bicchieri’s (Bicchieri 2006) gametheoretic formalization of social norms and Herbert Gintis’ (2009) definition of norms as
correlated devices. The sixth section explains why these accounts necessarily lead us toward a
non-individualistic understanding of game theory. The seventh section briefly concludes.
2. Constitutive Rules in Social Sciences and Philosophy
What is it to play baseball, to play chess, to break the law or to promise something? All these
activities share the common property that they cannot be described without referring to one or
several rules. This makes these rules somewhat special: they define a practice in the sense that
the practice could not exist without them. In philosophy, this kind of rules is generally called
constitutive rules because they are constitutive of a practice or an activity. This section
describes how some prominent philosophers have conceptualized constitutive rules and
indicates why to account for these rules is important for social scientists and in particular for
economists.
The concept of constitutive rules is essentially a product of the 20th century philosophy. Its
first obvious discussion is to be found in the later writings of Ludwig Wittgenstein, especially
in his Philosophical Investigations (1953). In this work, Wittgenstein developed – but never
precisely defined – the concept of language-game whose function was to capture the logic of
the ordinary and everyday use of language. As this is well known, the concept builds on an
analogy with the rules of games: speaking in a language is like playing in a game; it refers to a
bunch of activities and practices whose meaning is attached to a set of rules making these
3
very activities and practices possible. As a result, Wittgenstein’s argument was that a word
has no determinate meaning beyond the one attached to its everyday use by a community of
speakers. Language is fundamentally social because it is the product of collective practices.
To speak a language is like to play a game according to its rules: the fact that one is speaking
a language  consists in nothing more than the fact that one is following a set of rules
corresponding to .3
A few years later, John Rawls developed a similar account in his article “Two Concepts of
Rules” (Rawls 1955). Rawls’ discussion is focused on the issue of the relevance of the
distinction between the fact of justifying a practice and the fact of justifying a particular
action falling under the practice. Though Rawls’ made this distinction in the context of a
discussion of utilitarianism, it can be obviously extended beyond this restrictive scope. In
particular, Rawls discusses two conceptions of rules, only one of them being able to make
sense of the above distinction. According to the summary view, “rules are pictured as
summaries of past decisions arrived at by the direct application of the utilitarian principle to
particular cases” (Rawls 1955, p.19). Ignoring the reference to utilitarianism, the summary
view merely sees rules as behavioral regularities emerging from the application of a specific
principle of action to particular decision problems. Rawls gives a sensible illustration (1955,
p.22, emphasis are mine):
“suppose that a person is trying to decide whether to tell someone who is fatally ill what that
illness is when he has been asked to do so. Suppose the person to reflect and then decide, on
utilitarian grounds, that he should not answer truthfully; and suppose that on the basis of this
and other like occasions he formulates a rule to the effect that when asked by someone fatally ill
what his illness is, one should not tell him. The point to notice is that someone’s being fatally ill
and asking what his illness is, and someone’s telling him, are things that can be described as
such whether or not there is this rule. The performance of the action to which the rule refers
doesn’t require the stage-setting of a practice of which this rule is a part.”
As it will appear clearly later, Rawls’ summary view of rules has a strong affinity with the
standard account of rules-as-behavioral-regularities that permeates game theory. The point is
that the rule functions only as a description of the behavior observed in a series of specific and
similar cases. Therefore, particular cases are logically prior to rules (Rawls 1955, p.23).
Rawls contrasts the summary view with what he calls the practice conception: “[o]n this view
rules are pictured as defining a practice” (Rawls 1955, p.24). In this conception, “rules of
practice” are logically prior to particular cases where one’s behavior instantiates a practice:
the fact that one’s behavior in a particular context is associated to a specific practice depends
on the fact that one’s behavior conforms to the rules defining this very practice. A corollary is
then that for one to being taught a practice, one has to learn the rules defining it. Rawls
illustrates this conception with the game of baseball: playing baseball implies many basic
actions such as throwing a ball, running, sliding or swinging a piece wood having a specific
shape. But a set of related actions are possible and make sense only within the rules of
3
Some philosophers and social scientists have argued that Wittgenstein’s accounts of language games can be
extended to the general phenomenon of rule following. See for instance the sociologist David Bloor (1997) who
argues that one can find in Wittgenstein’s later writings the seeds of a theory of social rules and institutions.
4
baseball: hitting a home-run, stealing a base, striking out the batter, catching a fly ball, and so
on: “[n]o matter what a person did, what he did would not be described as stealing a base or
striking out or drawing a walk unless he could also be described as playing baseball, and for
him to be doing this presupposes the rule-like practice which constitutes the game” (Rawls
1955, p.25). The practice of playing baseball, which is defined by the fact of following a set
of well-defined rules, is thus logically prior to any specific case where the practice is
instantiated. The point of the practice conception rules is that it is able to make sense of the
distinction between the fact of justifying a practice and the fact of justifying a particular
action taking place within a practice. For instance, the fact that I am running after having hit
the ball is justified by the fact that I am playing baseball and thus following some particular
rules. To argue that this action is not well-founded would be to argue against the rules of
baseball. Now, one may argue against the relevance of playing baseball, but again it would
consist in arguing against the rules of baseball. In a nutshell, it is impossible to talk about the
fact of playing baseball without making reference to the rules of baseball.
More recently, the American philosopher John Searle has forcefully argued for a distinction
between regulative rules and constitutive rules as a foundation for a theory of institutional
facts ((Searle 1995); (Searle 2010)). The former regulate an existing practice or activity.
Several regulative rules are compatible with a specific practice. A change of regulative rule
would not alter significantly the practice. For instance, the activity of driving is not
fundamentally changed if a rule prescribing to drive at the left side of the road is substituted
for a rule prescribing to drive at the right side of the road. By contrast, constitutive rules
cannot be distinguished from the practice they constitute. According to Searle, they have a
particular form of the kind “X counts as Y in C”, where X and Y refer to two physical or
institutional objects and C to a set of circumstances where the formulation applies. Basically,
a constitutive rule grants to an object or to a state of affair a peculiar status which is
collectively recognized by the members of a population. For example, when one is moving
one square forward is pawn during a chess match, this move (a physical fact) counts as a
pawn move according to the rules of chess. This (institutional) fact is established as soon as
the move is made in the proper circumstances, in particular if at least two persons (the
players) recognize that they are playing chess. Similarly, the very possibility to checkmate his
opponent (a state of affairs) presupposes that the players recognize that they are following a
set of rules defining chess.
Searle’s constitutive rules have some interesting properties: they are performative and they
create deontic powers. Constitutive rules literally create a new institutional reality: they make
something as a case by representing it as being the case (Searle 2010, p.97). For instance, if a
group of people collectively agree that such and such pieces of paper count as “money”, then
this agreement creates money. It follows that constitutive rules create deontic powers since
any assignment of a function Y to a given object or person X consists in recognizing that X
must/can/cannot/may do or allow something. These deontic powers are themselves
constitutive of the function Y: to grant the status of “money” to a piece of paper implies to
recognizing that this piece of paper has, say, the power to purchase goods.
Though Wittgenstein’s, Rawls’ and Searle’s respective definitions of constitutive rules do not
fully overlap, they share a central thread. It is worth to make it explicit since that will help to
5
understand why an account of constitutive rules in game theory is needed. Most rules merely
regulate preexisting practices that, at least in principle, could exist without them. These
regulative rules may facilitate a social activity or even contribute to enhance its efficiency, but
the same social activity may be effective without those rules. This is not true in the case of
constitutive rules. Basically, to say that a rule R is constitutive of a practice P is to say that the
very fact of P-ing is to follow R. Thus, it is impossible to account for or even to describe P
without referring to R. Hence, to distinguish a constitutive rule from a regulative rule I
propose the following counterfactual reasoning:
Consider proposition : “for any actual world @ where the practice P is effective and
agents follow a rule or a set of rules R, there is a possible world W where P is effective
and agents follow a different rule or set of rules R*”.
If  is true, then R is/are regulative; if  is false, then R is/are constitutive.
If one accepts this definition of constitutive rules, then it is not difficult to see why social
scientists including economists should be interested in accounting for them. Assuming that
some significant socioeconomic practices are defined by constitutive rules, it is impossible to
explain these practices without explaining the relevant constitutive rules. Of course, the key
point is that since constitutive rules are prior to the practices they define, the social scientist
cannot explain them as being the product of independent goals or beliefs motivating the
agents. As we shall see, the standard account of rules in game theory is ill-defined to study
constitutive rules precisely because it defines rules as behavioral regularities. Other gametheoretic approaches are available though.
3. The Standard Game-Theoretic Account: Rules as Behavioral Regularities
As I have indicated in the introduction, game theory has been used for at least two decades by
economists to study several issues related to social rules: how they emerge and evolve, what
are their functions and how they affect people’s behavior. David Lewis’s pioneering book on
conventions (Lewis 1969) is the first major attempt to build a theory of social rules (more
specifically, of conventions) through game theory. Since, there have been many attempts to
account for the emergence and the working of social rules with game theory. 4 These attempts
mostly fall in what I call the standard game-theoretic account of rules, or Standard Account
for short. They share at least three characteristics: firstly, they are mainly concerned either (or
both) with explaining how rational agents are able to cooperate in spite of the fact they their
interests are not perfectly aligned or with how rational agents are able to coordinate in spite of
the indeterminacy resulting from the multiplicity of coordination equilibria. Secondly, they
tend to adopt a mostly evolutionary point of view: their main concern is to explain how rules
have emerged and evolved through the repeated interactions of agents endowed with various
forms of rationality. Thirdly, they conceive rules essentially as behavioral regularities
4
A non-exhaustive list would include Ullmann-Margalit (1977), Taylor (1987), Schotter (1981), Sugden (1986),
Young (1998).
6
resulting from the decentralized choices of the players in the game. While the first
characteristic is not specific to the Standard Account, the second and the third are.
Incidentally, they are the main reasons why the Standard Account cannot properly capture the
specificity of constitutive rules.
The essence of the Standard Account is well summarized by a recent article of Philip Smit et
al. (2011, p.7, emphasis in original):5
“We claim that all institutional facts are essentially tied to a set of actions. Traffic lights are
objects that I am incentivized to stop at, etc.; borders are things I am incentivized not to cross;
money is something I am incentivized to acquire for exchange, etc. We do not think that
institutional objects give rise to these actions and incentives; rather, we think that they simply
are natural objects individuated by these actions and incentives”.
Smit et al. present the Standard Account in the context of a critique of Searle’s theory of
institutional facts (hence the focus on “institutional facts”) but it is rather easy to generalize:
what they call “institutional objects” are indeed instances of social rules that agents are
following. Crucially however, the rules do not give rise to their behavior. Instead, the
institutional objects are the product of these behaviors. Put another way, according to Smit et
al., social rules are singled out because of behavioral regularities that make them salient. The
explanation of these behavioral regularities is to be found in the incentives leading people to
act in some way rather than in another. Therefore, one can say that in this standard account
rules are merely epiphenomenal: they emerge from the agents’ actions but do not play any
causal or functional role in the rise of these actions. As a consequence, they also cannot be
part of any explanation of these actions.
As argued above, the Standard Account is “evolutionary” in a loose sense of the term: rules
are behavioral regularities that arise spontaneously from various kinds of evolutionary
mechanisms involving adaptive play and/or learning; it is committed to the production of
“invisible hand explanations” (Aydinonat 2009). The defining property of social rules is that
they correspond to self-enforcing behavioral patterns: given the history of past plays in a
game, each agent is incentivized to reproduce the behavioral pattern, thus leading to a further
reinforcement of the pattern. As an instance of this evolutionary and behavioral approach to
rules, consider Andrew Schotter’s theory of social institutions (Schotter 1981). Schotter
informally defines social institutions as “regularities in behavior which are agreed to by all
members of a society and which specify behavior in specific recurrent situations” (Schotter
1981, p.9). More formally,
“A regularity R in the behavior of members of a population P when they are agents in a
recurrent situation  is an institution if and only if it is true that and is common knowledge in P
that (1) everyone conforms to R; (2) everyone expects everyone else to conform to R; and (3)
either everyone prefers to conform to R on the condition that the others do, if  is a coordination
problem, in which case uniform conformity to R is a coordination equilibrium; or (4) if anyone
ever deviates from R it is known that some or all of the others will also deviate and the payoffs
5
I have proposed a critical discussion of the account of Smit et al. relatively to their critique of John Searle’s
theory of institutional facts in Hédoin (2013a).
7
associated with recurrent play of  using these deviating strategies are worse for all agents than
the payoff associated with R” (Schotter 1981, p.11, emphasis in original).
Schotter focuses on two kinds of strategic interactions: coordination problems and prisoners’
dilemma problems (see figures 1 and 2):
Figure 1: Coordination problem
Player 2
Player
1
C
D
C
1;1
0;0
D
0;0
1;1
Figure 2: Prisoners’ Dilemma problem
Player 2
Player
1
C
D
C
2;2
0;3
D
3;0
1;1
The games depicted in figures 1 and 2 exemplify two different problems: in the coordination
problem, the difficulty lies in the multiplicity of Nash equilibria which makes impossible for
the game theorist to predict and to explain the players’ choices. In the prisoners’ dilemma
problem, the difficulty comes from the fact that cooperation (strategy C) is a dominated
strategy leading as a result even minimally rational players to defect in spite of the fact that
cooperation is socially preferable. However, ultimately these two distinct problems subsume
into a sole one because Schotter is not interested in the one-shot version of these games but
rather in their indefinitely iterated version. Hence, he defines L as the supergame consisting of
a basic game  depicting either a coordination problem or a prisoners’ dilemma problem and
which is played indefinitely or infinitely between a population of at least two players who
have perfect information of the history of past plays. A famous set of results sometimes called
“the folk theorems of repeated game theory” indicate that if the players’ discount rates are not
too high, then there are an infinite number of equilibria guaranteeing to each player at least his
minimax gain. This result comes from the fact that once a game is infinitely repeated,
conditional strategies become available. For instance, assume that in the prisoners’ dilemma
player 1 is playing a conditional strategy known as “grim trigger” (henceforth denoted GT)
which consists in cooperating at the beginning of the game and until the opponent defects, and
then to defect unconditionally. Then, if we denote 2 = 1/(1+r2) the discount factor through
8
which player 2 evaluates the present value of his future gains, player 2 is incentivized to
reciprocate cooperation if he cannot make a higher gain by defecting. With the payoff matrix
of figure 2, that means
2(GT ; GT) = 2/(1 - 2) > 2(GT ; D) = 3 + 2/(1 - 2)
2 > ½
Since an infinite number of such conditional strategies are available, the supergame L has a
large number of equilibria. As a result, the coordinate problem reappears since the players
must coordinate on a set of consistent conditional cooperating strategies if they do not want to
be caught in the defection equilibrium.6 In Schotter’s framework, the actual outcome on
which the players will coordinate is “selected” by a social institution. More exactly,
institutions depend on social norms defined as “informational devices that the agents of
societies develop to help them place subjective probability estimates over each other’s
actions” (Schotter 1981, p.52). Schotter argues that these norms emerge through a Bayesian
learning process where each player i deduces a mixed strategy si given a norm p summarizing
each player’s beliefs regarding the probabilities that the other players play their different
available pure strategies. More formally, in the case of a two-player game  corresponding to
a supergame L with n pure strategies available for both players, at any period t, there is a
norm p = (p1, p2), where p1 = (pi1,..., pin) is the vector of probabilities that player 2 associates
to each of the n strategy that player 1could play. Then, the pair p = (p1, p2) express the
subjective assessment of the type of behavior each player can expect from the other. Schotter
(1981, p.66) calls a “static Bayesian solution procedure” (SBSP) any function M that for a
given game L assigns a pair of mixed strategies s = (s1 ; s2) to each norm p = (p1, p2). Many
kinds of SBSPs are in principle available but a plausible restriction consists to requiring that if
p is a norm where each player assigns a probability 1 that the other will play his part in an
equilibrium s* = (s1* ; s2*), then both players will play their part in the equilibrium.7
From the function M , it is then possible to infer a norm updating procedure mapping any
pair pt = (p1t, p2t) at time t into a pair pt+1 = (p1t+1, p2t+1). Since the SBSP associates a unique
pair of mixed strategies s = (s1 ; s2) to any norm pt, the procedure determines a unique random
walk whose limiting behavior defines a stochastic process corresponding to a Markov chain.
This dynamic process thus makes the formation of beliefs entirely endogenous. More
importantly given the problem at stake here, the social rule (the norm p) reflects the totality of
the relevant information and is entirely defined by the game’s history. The norm is
institutionalized once the dynamic process reaches one of its absorbing states: once an
equilibrium point is reached, beliefs become self-fulfilling because players’ actual play
reinforces their expectations. We can denote p* = (p1*, p2*) such an institution or rule. Is p* a
constitutive or regulative rule? As noted above, the folk theorems imply that any supergame L
corresponding to an infinitely repeated basic game  has several absorbing states: even if we
6
As Schotter (1981, p.63) notes, the main difference is that in coordination games there is no unique policing
strategy, i.e. a strategy that one could play to sanction the other’s player deviation. But actually, because in
coordination games the players’ interests are aligned, no policing strategy is needed.
7
Obviously, this entails nothing more than assuming that players are minimally rational.
9
consider only norm updating processes whose SBSP satisfies minimal criteria of rationality,
the random walk may converge toward different equilibrium points p*. The random walk is
unique in the sense that the transition probabilities between any two states are known and
well-defined, but the actual dynamic is still indeterminate. As a consequence, several rules
may emerge on the basis of the same supergame L. This indeterminacy indicates that the rules
p* cannot be but regulative rules. They merely regulate the practice associated to L in that
they prescribe what the players should do through a vector of mixed strategies s*.8 But they
do not define or change the game properly speaking. For instance, consider an infinitely
repeated prisoners’ dilemma where, in addition to unconditional cooperation and
unconditional defection, the only other (conditional) strategy is grim trigger. With the payoff
of figure 2 and assuming a discount factor 1 = 2 = 0.9, we obtain the following supergame
(see figure 3):
Figure 3: The infinitely repeated Prisoners’ Dilemma with Grim Trigger
Player 2
C
C
Player 1 D
D
20 ; 20 0 ; 30
30 ; 0
GT
20 ; 20
10 ; 10 12 ; 10
GT 20 ; 20 10 ; 12 20 ; 20
This game has two equilibria: the equilibrium of unconditional defection and the equilibrium
of conditional cooperation through the grim trigger strategy. The stochastic process defined
by the SBSP and the associated norm updating procedure can either converge toward one
equilibrium or the other. Granted we can set particular conditions upon the SBSP, it is
possible to determine the transition probabilities of the resulting Markov process and
therefore the respective probabilities of convergence toward the two equilibria. However, it
should be clear that whatever the actual dynamic is, the game (and the practice it depicts) is
still the same: using the counterfactual reasoning I presented in the preceding section, it is
easy to imagine two possible worlds (here, two games) closely identical except that in one the
players are cooperating while in the other they are defecting.
This result is not surprising because in Schotter’s model, as well as in all game-theoretic
models conforming to the Standard Account, rules are merely epiphenomenal. In conformity
with the quote of Smit et al. above, the rules defined by the vectors of vectors probabilities p
do not give rise to the action of the players. They merely summarize what has happened
during the past plays of the game. Strictly speaking, since what we have is a Markovian
process, what determines the present play of the agents is the present state (as well as their
learning rule) and since the game allows for indetermination and randomness, there is no one-
8
Note however that in Schotter’s model, a rule p* is not sufficient to prescribe a particular profile of actions for
the player. The logical inference of s* from p* depends on the particular SBSP the players are using.
10
to-one relationship between the game and the outcome.9 The parallel between the standard
account of rules in game theory and what Rawls called the “summary view” appears clearly:
the performance of the action by the agents (i.e. their actual behavior in the game) is not
determine by the setup of any prior practice; the rules do not define the game, they merely
describe what the agents have done in the game.
4. Constitutive Rules and Psychological/Behavioral Game Theory
In this section, I present a first game-theoretic attempt that breaks with the Standard Account
and which is potentially relevant regarding the issue of constitutive rules. I associate this
attempt to psychological/behavioral game theory (respectively PGT and BGT).10 PGT and
BGT have been intended as theoretical and methodological attempts to capture what people
are actually doing in strategic interactions, rather than making normative or prescriptive
claims of what rational agents ought to do. The main motivation for the development of both
PGT and BGT consists in a twofold fact. On the one hand, traditional game theory leads to
some recommendations that at best are hardly implementable, at worst absurd. The backward
induction principle is a great example of this problem. On the other hand, it is now well
acknowledged that individuals do not behave as game theory predicts or recommends, either
in experimental setups or in real interactions. Though the reasons for these latter deviations
are manifold, a significant one is that people actually do have “non-standard” preferences
which in particular depend on the beliefs they entertain regarding others’ intentions and
beliefs. As I will show below, it is in this context that constitutive rules may be introduced.
Matthew Rabin’s article “Incorporating Fairness into Game Theory and Economics” (Rabin
1993) has been a major step in incorporating non-standard preferences in a game-theoretic
framework. Rabin’s model demonstrates that it is possible to explain some “surprising”
phenomena (such as the fact that individuals cooperate in a one-shot prisoners’ dilemma) if
agents have a preference for fairness and (positive and negative) reciprocity. More exactly, if
one expects the other player to be “kind”, then ceteris paribus one will prefer to reciprocate
this kindness by behaving in a way that the other player will also consider as “kind”. Kindness
is identified by the fact that one sacrifices intentionally a potential material gain such as the
other player can increase his own material gains relatively to what he is “entitled to”. In a
prisoners’ dilemma, kindness is then simply associated to cooperation: considering that the
payoff resulting from mutual defection is what the players are entitled to,11 if player 1 expects
9
One may argue that since once we know the specific properties of the SBSP it becomes possible to determinate
precisely the transition probabilities of the resulting Markov chain, there is no longer indetermination. I think
this is a valid argument but not a refutation of my claim since in this case this is not the rule that breaks
indetermination, but the properties of the SBSP.
10
These two terms are not fully synonymous. Following the pioneering work of John Geanakoplos et al. (1989),
a psychological game is a game where each player’s payoff depends not only on the other players’ actions but
also on the set of iterated beliefs regarding what everyone believes and does. Payoffs are thus beliefs-dependent.
Though there is no “official” definition for behavioral game theory (but see (Camerer 2003)), it seems sensible
to refer to it as the use of a set of experimental and empirical methods to test and advance traditional game
theory. Therefore, PGT refers to a particular theoretical framework, while BGT rather refers to a methodological
approach to game-theoretic problems. Obviously, there are significant overlapping areas between the two.
11
Actually, in Rabin’s model, players’ entitlement are computed slightly differently since the player’s i
entitlement when he plays strategy si corresponds to the average material gains related to the non dominated
11
player 2 to cooperate, then it follows that player 1 should consider player 2 as “kind”. If
player 1’s preference for fairness and reciprocity is sufficiently strong relatively to his
preference for material gains, then he will reciprocate player 2’s kindness by cooperating. In
this case, mutual cooperation can be a fairness equilibrium, i.e. a Nash equilibrium where
fairness is incorporated into the players’ utility functions.12
Robert Sugden’s (2000) game-theoretic model of resentment aversion builds on a similar
logic. Starting from Lewis’ theory of conventions (Lewis 1969), Sugden argues that
established conventions tend to acquire a normative force through time and ultimately become
social norms. If initially conventions hold because it is in everyone interest (narrowly defined)
to play his part in an equilibrium outcome if each expects others to do so, the motivation to
conform to the behavioral regularities is progressively reinforced further by the fact that
expectations become normative: “Roughly, my claim is that it is a property of human
psychology that we feel a sense of resentment against people who frustrate our reasonable
expectations, and also we feel aversion towards performing actions which are likely to arouse
people’s resentment” (Sugden 2000, p.112). Once resentment aversion is incorporated in the
players’ utility functions, a very similar mechanism to the one exemplified by Rabin’s model
is activated: basically, if I expect for whatever reason my opponent/partner in a game to
expect me to play a particular strategy si, then I know that by playing any other strategy si’ I
will frustrate his expectations. If I consider my opponent/partner’s expectations to be
legitimate13, then I know he will rightly feel resentment toward me; as a consequence, it will
diminish my utility. What distinguishes Sugden’s model from Rabin’s model is the interesting
fact that in the former expectations are generated by a preexisting norm while in the latter
they do not have well-specified origins. Moreover, the normative character of expectations in
Sugden’s model is tied to this preexisting norm. In Rabin’s model, expectations are not
normative; only intentions are subjected to a normative evaluation (“kind”, “unkind”) which
directly depends on the payoff matrix. Still, Sugden’s model leads to a very similar
conclusion: for instance, in a one-shot prisoners’ dilemma, assuming that the players’
resentment aversion is sufficiently strong, the fact that it is common belief or knowledge that
cooperation is the norm will lead the players to actually cooperate. Mutual cooperation
corresponds then to a normative expectations equilibrium.
What differentiate Rabin’s and Sugden’s models from the Standard Account is that because
payoffs are a function of the players’ beliefs (in particular their second-order beliefs), then the
very nature of the game changes as a function of the beliefs the players are actually
entertaining. If we assume, as Sugden does, that the players’ beliefs are formed on the basis of
rules such as a social norm, then we must conclude that a change in the norm may change the
game the agents are playing. In this sense, the rule becomes constitutive of the game. I will
make this crucial point clearer by using Cristina Bicchieri’s game-theoretic model of social
outcomes. As a consequence, when one cooperates in a prisoners’ dilemma, he is merely entitled the material
gains corresponding to the outcome where the other player defects.
12
Since a fairness equilibrium is a Nash equilibrium, the players’ beliefs must be consistent and thus be
confirmed by the maximizing behavior of the players.
13
In Sugden’s model, the legitimacy of expectations is guaranteed by the fact that the normatively appropriate
strategy profile in the game (i.e. what each player ought to play) is assumed to be common knowledge among the
players.
12
norms as a third example (Bicchieri 2006). Indeed, Bicchieri’s model is ultimately a powerful
game-theoretic account of constitutive rules.
According to Bicchieri, social norms are grounded on the conditional preference for
conformity owned by most agents. Social norms are related to practices that one conforms to
because he expects others to conform and he believes that others expect him to conform.
Bicchieri’s accurate definition is the following (Bicchieri 2006, p.11):14
A behavioral rule R for situations of type S, where S corresponds to a mixed-motive
game, is a social in a population P if there exists a sufficiently large subset Pcf  P, such
that for each agent i  Pcf:
(1) i knows that a rule R exists and applies to situations of type S (contingency
condition);
(2) i prefers to conform to R in situations of type S on the condition that (conditional
preference condition):
(a) i believes that a sufficiently large subset Pf  Pcf conforms to R in situations of
type S (empirical expectations);
and either
(b) i believes that a sufficiently large subset of Pf  Pcf expects i to conform to R in
situations of type S (normative expectations);
or
(b’) i believes that a sufficiently large subset of Pf  Pcf expects i to conform to R in
situations of type S, prefers i to conform, and may sanction behavior (normative
expectations with sanctions).
A social norm is followed if the subset Pf  Pcf is sufficiently large such that for all agents i 
Pf condition (2a) and either condition (2b) or (2b’) are met. Note that Pcf corresponds to the
set of conditional followers of the norms, i.e. those individuals that know R and prefer to
conform to R conditional on the fact that a sufficient number of other persons also conform to
R; Pf corresponds to the set of actual followers. The fact that Pcf  P and Pf  Pcf indicates
that a norm can be effective in a population P even if some individuals do not care about
following the norm and if a set of individuals do not actually follow the norm.
As the above definition makes it clear, rule followers must have a conditional preference for
conformity. This preference can be easily introduced into a proper model with an augmented
utility function (see Bicchieri 2006, p.52-54):
(1)

14
Bicchieri distinguishes social norms, moral norms, personal norms and conventions. Only social norms and
conventions give expectations an important function (the fact of following a moral norm or a personal norm do
not depend on what others are doing or expecting) and only in the case of social norms expectations do have a
normative character. Only social norms (and maybe moral norms) can be said to be constitutive in Bicchieri’s
theory.
13
This equation states that i’s utility associated with playing his part in a strategy profile s = (s1,
..., sn) depends on two factors: firstly, the material payoff i as it is traditionally defined in
game theory; secondly, the payoff associated to the fact of either a) conforming to a norm Nj
when everyone conforms, b) conforming when at least one person do not conform, c) of not
conforming when everyone else conforms, or d) of not conforming when nobody else
conform. Equation (1) indicates that if any player j (including i) deviates from the norm, then
all the players m suffer a psychological loss corresponding to the difference between what
they would have gain provided everyone would have conformed and what they have actually
gained. The parameter ki  0 is thus a constant reflecting the importance each agent attaches
to conforming to the norm and to others’ expectations.
This utility function gives rise to interesting complications. In particular, as Bicchieri
demonstrates, the fact that a social norm exists may entail a transformation of the game the
agents are playing, provided that the latter’s conditional preference for conformity is
sufficiently strong. More specifically, the property of a social norm is to transform a mixedmotive game where the interests of the players are partially conflicting into a coordination
game.15 Take the two-person, one-shot prisoners’ dilemma of figure 2 above and assume that
player 1 attaches some importance to conforming to the norm of cooperation that is deemed to
be effective between him and player 2, i.e. k1 > 0. Moreover, assume also that k2 > 0 and that
k1 = k2 > 0 is common knowledge among the players. It is easy to check that the game is no
longer a prisoners’ dilemma (see figure 4):
Figure 4:
The Prisoners’ Dilemma with common knowledge preference for conformity
Player 2
Player
1
C
D
C
2;2
-2k ; 3 – 2k
D
3 – 2k ; -2k
1;1
If k > ½, defection is no longer a dominant strategy for both players. Since their conditional
preference for conformity is assumed here to be common knowledge, they have common
knowledge that they are actually playing a coordination game where mutual cooperation is
the Pareto-optimal outcome. Now, assume that player 1 has a conditional preference for
conformity but does not know whether k2 > ½ or not. Then, the game takes the form of a
Bayesian game where player 1 attaches a probability p of playing a coordination game (if k2 >
½) and a probability 1 – p of playing a prisoners’ dilemma (if k2 < ½). His actual choice will
15
It follows from this point that social norms do not operate in what are “true” coordination games. Therefore,
according to Bicchieri, constitutive rules do not play a role in coordination problem. This point may be disputed
on the ground that some conventions may be constitutive rather than merely regulative (Marmor 2009).
14
thus depend on the value of p. Clearly, the stronger the norm in the population, the higher p
will be.
The above developments clearly indicate that social norms in Bicchieri’s game-theoretic
model are constitutive rules. Basically, the nature of the practice depicted by the game (i.e.
whether the game is a coordination game, a mixed-motive game or a Bayesian game) depends
on whether a social norm exists and is actually followed by the members of the population. As
in Sugden’s model of resentment aversion, the payoff matrix is a function of the players’
beliefs and those beliefs are determined by the social norm. If the norm changes, the game
then necessarily changes. This point is reinforced by the fact that Bicchieri explicitly rejects
the Standard Account: “A norm cannot be simply identified with a recurrent, collective
behavioral pattern ... Avoiding a purely behavioral account means focusing on the role
expectations play in supporting those kinds of collective behaviors that we take to be normdriven” (Bicchieri 2006, p.8-9). As the last quote indicates, the constitutive nature of rules in
PGT/BGT comes from the particular treatment of expectations: expectations do not merely
combine with preferences to lead one to make a choice; they also change preferences. In this
case, rules are not related to behavioral regularities but to sets of iterated (first and secondorder) beliefs that contribute to define the games the players are playing.
5. Constitutive Rules and Correlated Equilibria
Imagine you own stocks and that you are uncertain regarding how their prices will evolve.
Then, on television you see the Chairman of the Federal Reserve announcing that the financial
sector is on the verge of a meltdown. Following this announce, you expect the prices of your
stocks to fall fairly quickly since you know that the other agents on the financial markets also
heard the announce and will probably sell their stocks as a result. Therefore, you decide that
you should also sell your stocks. If everyone is reasoning the same way, then the prices will
actually fall and a meltdown will occur.16 Beyond the fact that what we have here is an
instance of a self-fulfilling prophecy, the interesting point is the role played by the Fed’s
Chairman who acts as a “choreographer” sending signals to each agent regarding the state of
the world. Formally, the Fed’s Chairman is a correlated device implementing a correlated
equilibrium.
In this section, building on the work of Robert Aumann (1987) and Herbert Gintis ((2009);
(2010)), I will show that social norms may be interpreted as implementing correlated
equilibria. I also show that as a consequence, they have the status of constitutive rules. First, I
start with some definitions. Informally, a correlated device is an entity (an agent or an object)
which observes the actual state of the world and then sends a signal to each player under the
form of an instruction indicating which strategy should be played. Each player knows his own
instruction but not necessarily the instructions sent to the other players. A correlated
equilibrium is then a situation where everyone follows the instructions sent by the correlated
device and where nobody has an interest to change his strategy given that others follow their
instructions.
16
This example is from David Levine (Levine 2011). Note however that he does not interpret this example as a
case of correlated equilibrium.
15
A formal characterization of the correlated equilibrium concept can be given using epistemic
game theory (see Gintis 2009, chap.7). An epistemic game G is a normal form game with a
set N = (1, ..., n) of n players, a set S = (S1 ... Sn) of pure strategies and a function ui: S  R
mapping any combination of strategies into a n-tuple of real numbers. The epistemic
components of the game include a set  of possible states of the world , a possibility set Pi
for each player and a subjective prior pi(. ; ). Each state of the world  specifies completely
all the relevant parameters that may be the object of uncertainty for the players in G (Aumann
1987, p.6). This includes in particular a complete specification of the strategy profile played
by the players at the state  and possibly a specification of each player’s conjectures at 
regarding how the game will be played. A possibility set Pi partitions  into units of
knowledge and defines a knowledge partition. It indicates what each player i knows and
believes when the actual state of the world is . Finally, the subjective prior pi(. ; ) is a
probability measure for each player i over . It specifies each player’s prior belief over the
realization of a particular state . Hence, a player’s subjective prior defines his beliefs over
the strategy profile that will be played by others. A correlated strategy in G is a function f
defined over a probability space (, p) and mapping  into S. A correlated strategy defines for
any random variable  an instruction fi()  Si for each player i. Denote (f-i, gi) the outcome
where everyone but one player i follows the instruction. The correlated strategy f forms a
correlated equilibrium if the expected gain for each player i resulting from following f is equal
or superior to the expected gain associated to the outcome (f-i, gi) for any random variable .
Aumann (1987, p.7-8) demonstrates that if we assume that a) players are Bayesian rational at
each state , b) players have a common prior p1 = ... = pn over , and c) each player chooses
a pure strategy si()  Si at , then the n-tuple s = (s1, ..., sn) corresponding to the distribution
of actions forms a correlated equilibrium defined by a correlated device f over a probability
space (, p), with f() = s() for all . The converse is also true: for any correlated
equilibrium distribution f in a game G, it is possible to define an epistemic game G with a
common prior p such that at every state , a Bayesian rational player will implement a move
si() corresponding to the correlated equilibrium (Gintis 2009, p.136). This result is important
because it shows that the existence of a correlated equilibrium is tied to the assumption of a
common prior over .17 Indeed, as Aumann himself recognizes, if we take this assumption for
granted, then the above theorem appears to be rather trivial: if all the (Bayesian rational)
players share the same prior beliefs over what will happen, then they cannot form inconsistent
conjectures if they are submitted to the same information.18 The same is true even if the
players’ respective information partitions differ in such a way they do not have the same
knowledge: as long as all players “know” the information partitions of the other players,19
17
See Morris (1995) for an insightful discussion of the common prior assumption in game theory and in
economics.
18
Aumann (1987) calls this idea the “Harsanyi doctrine” following John Harsanyi’s claim that rational agents
submitted to the same information must agree on the same conclusion. As Aumann (1976) has demonstrated in
another famous theorem, if rational agents have a common prior and if their posterior conjectures are common
knowledge, then they cannot “agree to disagree”.
19
Note that this is implicitly assumed in the epistemic model above. Since any state  is a complete
specifications of all the relevant information, at  any player knows the information partitions of the other
16
each player must know the other players’ conjectures and rationally update his own ones
accordingly. However, it should be clear that the common prior assumption is problematic in
most of the relevant socioeconomic cases where objective probabilities cannot be defined
and/or the definition of the state space  is ambiguous. While it is reasonable to expect that
we all the same prior regarding the probability that a coin falls on tail, it is hard to see why we
should all have the same prior belief regarding the probability that, for instance, a financial
meltdown occurs this year.
This objection is well-founded but a possible answer precisely lies in the existence of social
norms (Gintis 2009). Indeed, one may argue that social norms play the role of correlated
devices because they have the property to generate a common prior over a well-defined state
space. Crucially, it is in this context that social norms appear to be constitutive rules. I shall
make my argument by working through the example of what I call the “Property Game”
(figure 5):
Figure 5: The Property Game
Player 2
Player
1
C
D
C
2;2
0;4
D
4;0
-1 ; -1
The Property Game (also known as the “chicken game”) depicts a situation where two players
fight for an asset or a prize. A player may cooperate and propose to share the asset or defect
by trying to keep the asset for himself. If both defect, a fight follows and both players are hurt.
This game has three Nash equilibria: two in pure strategies where one player cooperates and
the other defects, one in mixed strategy where each player cooperates with probability 1/3.
Interestingly, while the pure strategy equilibria are highly unfair, the mixed strategy
equilibrium is inefficient since mutual defection occurs with probability 4/9. However,
assume that it is possible to distinguish between an “incumbent” and a “challenger” in any
occurrence of this game. Contrary to the challenger, the incumbent is the actual possessor of
the asset. Thus, the state space  contains two states 12 and 21 where one of the players is
identified as the incumbent (respectively player 1 and player 2). To keep things simple, I
assume that the respective role of the two players is always non-ambiguous: not only each
player knows for sure whether he is the incumbent or the challenger20 but also each player
knows that, knows that the other knows that, and so on. Denote E any event revealing one of
the two states. By definition, E is a public event and is thus common knowledge. Correlated
players. The same is true for the prior p: at each , any player i knows the prior pj of any player j. By the same
reasoning, all players know the preceding. This implies that in an epistemic game G, the players have common
knowledge of their information partitions and their priors (Aumann 1987, p.9). In the contrary case, that would
mean that the description of  is incomplete and  should be split into two or more refined states ’.
20
Formally, that means that the possibility sets of each player each contains one and only one set, i.e. Pi  = 
for all i and all .
17
strategies mapping a state of the world into a specific pure strategy are now possible. In
particular, a norm N may specify that the incumbent should defect, while the challenger
should cooperate. Formally, for any players i, j, and for any event E, the norm N(E):   S
gives the following instructions to the player i:
si = D, if ij
si = C, if ji
Since any state  completely specifies the strategy profile implemented by the players, and
since the events E are common knowledge, each player makes a sure conjecture regarding
what the other will do at any . If i is instructed by N to play C, he knows that the actual state
of the world is ji. Thus, he knows that j will play D; clearly, i does not have any incentive to
deviate from the instruction given by N. Since the same is true at ij, N(E) is a correlated
equilibrium. Interestingly, in this particular example, at the equilibrium, N(E) is also common
knowledge because each player is able to deduce this reasoning by himself, knows that the
other is also able to do, knows that the other knows that, and so on.21 Taking some liberties
with the notation, denote CK(X) for “X is common knowledge among the players” where CK
is a knowledge operator with the traditional properties of S5 modal logic. Then, what
happened in the Property Game can be described as
CK(E)  CK(N)
This expression depicts an indication relation: a public event E indicates a norm N to all the
players and this is commonly known. The notion of indication relation is to be found in
Lewis’ theory of conventions ((Lewis 1969); (Cubitt & Sugden 2003)) and is tightly related to
the assumption that players are symmetric reasoners, i.e. they make the same inferences on
the basis of the same facts or propositions ((Gintis 2009, p.142); (Vanderschraaf 1998)).
Actually, and unsurprisingly, the common prior assumption and the symmetric reasoning
assumption are formally equivalent (Hédoin 2013b). Therefore, it follows that if the norm
N(E) is a correlated equilibrium as I have shown above, that implies that it also instantiates a
common prior p(.; ) over a state space .
This epistemic model makes a strong claim for the constitutive nature of the norm N(E).
Indeed, the norm N(E) relies on the practical inference associating a strategy profile s*() to
an event E. The norm builds on and at the same time creates the possibility to make this
inference. In this sense, the norm and the related practical reasoning of the form E  s*()
are constitutive of each other. Moreover, the strategy profile s*() implemented by the
players is logically related to a belief profile *() = (1, ..., n) where i is player i’s
21
As noted above, this is not generally true. If if information is imperfect, the players do not necessarily know
what is the actual state . Hence, they may not know what the other players are actually doing. However, each
player is able to form a set of conjectures at  and by definition (since information partitions and priors are
common knowledge, see footnote 19 above) they are common knowledge. It can be seen thus that a correlated
equilibrium is a specific type of mixed equilibrium strategy where the players do not explicitly randomized.
Again, this is noted by Aumann (1987).
18
conjectures at  upon which he maximizes expected utility. It follows that the norm is also
related to an epistemic reasoning of the form E  *(). Now, analogically to the
Wittgensteinian question “what is the fact of following of adding two numbers?” (Bloor
1997), we could ask “what is the fact of playing his part in s*()?”. Our answer in the
correlated equilibrium framework should be nearly identical to Wittgenstein’s. According to
Wittgenstein, there is nothing more in the fact of making an addition than the fact of
conforming to a particular practice where everyone collectively agrees that they are making
an addition (Bloor 1997, p.58-73). The rule of addition and the practice of making an addition
are mutually constitutive of each other: to make an addition is to apply the rule of addition
according to the practice of what is to make an addition in a particular community. This is
very similar to what happens in the Property Game: to follow the norm N(E) is nothing more
than using the practical and epistemic reasoning E  [s*() & *()]; but to reason
according to E  [s*() & *()] is the fact of following N(E). Thus, a social norm defined
as a correlated equilibrium is a constitutive rule.
6. Constitutive Rules and the Non-Individualistic Understanding of Game Theory
I have presented above three different game-theoretic approaches of rules: the “rules-asbehavioral-regularities” account, the “rules-as-normative-expectations” account and the
“rules-as-correlated-devices” account. I have suggested that rules are constitutive only in the
latter two, though for different reasons. In the “rules-as-normative-expectations” approach
exemplified by Bicchieri’s theory of social norms, social norms have the property to change
the game the agents are playing: they transform a mixed-motive game into a coordination
game. This ability comes from the fact that norms depend on a conditional preference for
conformity entering directly into the agents’ utility functions. As a consequence, their
preferences are a function of their beliefs, and hence of the norms. In the “rules-as-correlateddevices” approach, social norms are constitutive because they set the stage for the fact that the
agents have a common prior and lead them to behave in a particular way. In other words, the
norm defines a specific epistemic game and a change in the norm would change not only the
agents’ behavior but also the whole characteristics of the epistemic game (starting with the
agents’ prior).22
In this section, I shall point out that these two game-theoretic accounts of constitutive rules
dismiss any form of methodological individualism.23 This is not a contentious claim regarding
what can be called “strong” methodological individualism such as defined by Jon Elster
(1982, p.453): “the doctrine that all social phenomena (their structure and their change) are in
principle explicable only in terms of individuals – their properties, goals and beliefs”. Game
theory rejects this kind of methodological individualism since it explains social outcomes and
regularities through the interactions between rational agents endowed with preferences and
22
Note that the “rules-as-normative-expectations” account and the “rules-as-correlated-devices” approach are far
from being equivalent. For instance, Bicchieri (2010) notes that Gintis’ account of norms through the correlated
equilibrium concept is not able to explain how norms develop in games such as the prisoners’ dilemma. Indeed,
in a prisoners’ dilemma a rational player will never follow a “choreographer” instructing to cooperate.
23
This section expands previous arguments made elsewhere: Hédoin (2012), (2013a), (2013b).
19
beliefs. But most social scientists and virtually all game theorists endorse a weaker form of
methodological individualism, sometimes called “structural individualism”:
“In structural individualism, (...) actors are occupants of positions, and they enter relations that
depend upon these positions. The situations they face are interdependent, or functional related,
prior to any interaction. The result is a structural effect, as distinguished from a mere interaction
effect” (Udehn 2001, p.304).
My claim is that game theory, once it accounts for constitutive rules, no longer conforms even
to this weaker methodological individualism. To backup this claim, several points regarding
the social sciences in general and game theory in particular should be acknowledged. Firstly,
a traditional argument in defense of methodological individualism in the social sciences is that
social structures such as institutions are necessarily and only the product of the individuals’
choices and actions. As a consequence, it is at least in principle possible to reduce the
explanation of any social phenomenon to these choices and actions. However, this
“supervenience-implies-explanatory-reductionism” thesis that is also very popular in the
natural sciences24 fails at the epistemological level (Menzies & List 2010). Even if one
accepts the ontological postulate of reductionism (i.e. social structures are nothing but the
product of the interactions between agents), it does not logically follow that any relevant
causal explanation must proceed through reduction with the agents and their interactions as
the only explanans. Indeed, depending on one’s conception of causality, it can be perfectly
legitimate to consider that social structures are caused by other social structures. Secondly, it
can be argued that game-theoretic micro-explanations of institutions cannot work without
making some references to social structures (Hédoin 2012). The point is a consequence of the
indeterminacy problem: since multiple equilibria exist in most strategic interactions, social
structures are ultimately needed to introduce some determinacy on the model. This is
particularly clear when game theory is used in complement with a historical narrative: in this
case, history-dependent social structures break game-theoretic indeterminacy.25 Not only
institutions (such as social norms) may change the agent’s beliefs over the states of the world
and the actions of others, but they are also necessary to make sense of the agents’ beliefs and
behavior.
Thirdly, the epistemic foundations of game theory and of most of its solution concepts are
incompatible with methodological individualism, even in its weaker form. This point is wellput by Gintis (2009, p.162, emphasis in original):
“Epistemic game theory suggests that the conditions ensuring that individuals play a Nash
equilibrium are not limited to their personal characteristics but rather include their common
characteristics, in the form of common priors and common knowledge. We saw that both
individual characteristics and collective understandings, the latter being irreducible to individual
24
Actually, the argument has first been developed in philosophy of mind in the context of the mind/body
controversy. The “physical reductionism” thesis claims that since mental events and properties are necessarily
causally realized by physical events and properties located in the brain, then mental properties and events cannot
have causal power on their own. Therefore, any mental event should be entirely explainable by one or several
physical events.
25
For an illustrative example, see the insightful historical and institutional analysis of medieval economies
developed by Avner Greif (2006).
20
characteristics, are needed to explain common knowledge. It is for this reason that
methodological individualism is incorrect when applied to the analysis of social life”.
It is widely known that the conditions required for rational players to play a Nash equilibrium
are rather stringent, in particular with three or more players (Aumann & Brandenburger
1995). Not only players must have common knowledge of the game structure and of their
rationality, but they must also have common knowledge of their conjectures and common
priors. Arguably, all these conditions are seldom verified and more importantly they do not
reduce to individual properties. For instance, while rationality is a property owned by an
agent, common knowledge of rationality is an interactive property related to a population of
agents. The same is true for the common priors assumption which by definition is a collective
property: an agent cannot have “common priors” but a population of agents can. The point is
that both common knowledge (of rationality, of conjectures) and common priors are collective
properties which cannot be assumed to be necessarily owned by a population of agents. They
are rather the result of specific circumstances and of social, cultural and historical
mechanisms.
Acknowledging these points, it should be clear that the issue of constitutive rules only
contribute to make the inadequacies of methodological individualism stronger, in particular in
a game-theoretic framework. In one sense, this is not a surprise given the very definition of a
constitutive rule: a rule R that defines a practice P such that to P-ing is to follow R. It follows
from this definition that it is impossible to explain the behavior of agents in P without taking
into account the rule R. Therefore, R is an unavoidable component of the explanans regarding
the explanandum P. If game theory is able and must account for constitutive rules, as I have
shown in the preceding sections, then game theory is not and should not be tied to
methodological individualism.
It is useful to return once again on the reason why game theory, when it works as a theory of
constitutive rules, is not individualistic. Basically, this is because when one builds a gametheoretic model of a constitutive rule, the basic assumptions of the model regarding the
players’ preferences, knowledge, prior beliefs and available strategies depend on the rule.
This is true in the “rules-as-normative-expectations” approach where the rule directly enters
into the players’ beliefs and creates normative expectations. This is also the case in the “rulesas-correlated-devices” approach where the rule defines the players’ priors and makes them
consistent and identical. Now, compare with the Standard Account where rules are defined as
behavioral regularities. Game-theoretic models corresponding to this account build explicitly
or implicitly on a “state of nature” assumption, i.e. the starting point of the model is a world
without prior institutions or rules. An evolutionary or learning process then progressively
makes a behavioral pattern emerging. In this framework, the rule summarizes the behavioral
pattern but it should be clear that the rule never impacts the players’ beliefs or actions. Both
are not function of the rule but only of the beliefs and behavior of all other agents. The rule is
thus never constitutive because the model does not recognize anything beyond properties
related to the individual agents.
21
As a last note, one may object that the two game-theoretic accounts of constitutive rules
presented above lack a proper explanation of the evolution of rules, while the Standard
Account is able to explain how rules emerge and evolve. Actually, this is a good point and
one that indicates that the Standard Account is still valuable. However, it should be duly
acknowledged that ultimate explanations and proximate explanations are complementary and
not substitutes.26 To explain how a rule evolved is not sufficient to explain how it influences
the behavior of social agents. Moreover, regarding the evolution of rules and the distinction
between constitutive and regulative rules, two different claims could be made. It could be
argued that as far as the evolution of rules is concerned, the distinction between constitutive
and regulative rules is no longer relevant. Rules may become constitutive rules only after they
have emerged. What we would like to have then is an explanation of the mechanisms that lead
a rule to become constitutive rather than merely regulative. But it could also be argued, in
Searle’s fashion, that constitutive rules do not evolved but are rather created ex nihilo through
“declarations” (i.e. speech acts). The ability of humans to create constitutive rules through
speech acts is the product of our evolutionary history but the creation of constitutive rules is
purely a proximate phenomenon. For instance, one may be interested in the history of chess
and of its rules; but inquiring into the evolution of the rules of chess is of no help to explain
why when I sit in the front of a chessboard, I expect from my opponent a set of intentions,
beliefs and actions. The reason for these expectations is not the history of chess, but “simply”
the fact that I play “chess”, i.e. a well-defined practice among the members of a community.
If one accepts this argument, then an evolutionary explanation of constitutive rules seems
irrelevant.
7. Conclusion
Constitutive rules are an essential component of the social reality. They create the very
possibility to participate to fundamental practices such as playing games, selling and buying
goods, creating corporations, communicating, and so on. Since game theory has been heavily
used for the past three decades as a major tool in the elaboration of a theory of social rules,
one could ask whether game theory is able to account for this kind of rules. Moreover, game
theory needs constitutive rules for foundational reasons: indeed, constitutive rules are a way
to break the indeterminacy that surrounds most of game theory.
This article shows that actually game theory can account for constitutive rules but that this
entails giving up the individualistic understanding that permeates its use and its interpretation
by economists. Indeed, most game theorists study rules in the context of what I call the
Standard Account. The Standard Account corresponds to what Rawls dubbed the “summary
view”: rules are a mere summary of the history of the past plays in the game. In this
framework, rules are merely epiphenomenal and regulate agents’ interactions only indirectly
through the knowledge of the history of the game. There are approaches that break with the
26
The distinction between ultimate and proximate explanations comes from evolutionary biology. An
evolutionary explanation concerns the evolutionary mechanisms through which a trait or a behavior has evolved
(e.g. natural section, drift or social learning). A proximate explanation is about the biological or psychological
mechanisms that explains how the genetic, psychological or social features cause a particular behavior.
22
summary view of the standard account though: psychological game theory conceives rules
through the generation of normative expectations, and epistemic game theory conceives rules
as correlated devices. In both cases, I showed that the resulting account of rules is closer to
Rawls “practice conception” of rules. Rules are no longer epiphenomenal; they are an integral
component of the explanation of the phenomenon of interest.
References
Aumann, R.J., 1976. Agreeing to Disagree. The Annals of Statistics, 4(6), p.1236‑1239.
Aumann, R.J., 1987. Correlated Equilibrium as an Expression of Bayesian Rationality.
Econometrica, 55(1), p.1‑18.
Aumann, R. & Brandenburger, A., 1995. Epistemic Conditions for Nash Equilibrium.
Econometrica, 63(5), p.1161‑1180.
Aydinonat, N.E., 2009. The Invisible Hand in Economics: How Economists Explain
Unintended Social Consequences First Edition, Second Printing., Routledge.
Bicchieri, C., 1993. Rationality and Coordination, CUP Archive.
Bicchieri, C., 2006. The Grammar of Society: The Nature and Dynamics of Social Norms,
Cambridge University Press.
Bicchieri, C., 2009. Rationality and Indeterminacy. In D. Ross & H. Kincaid (Eds.), The
Oxford Handbook of Philosophy of Economics, Oxford University Press, p.159-188.
Bicchieri, C., 2010. Norms, preferences, and conditional behavior. Politics, Philosophy &
Economics, 9(3), p.297‑313.
Bloor, D., 1997. Wittgenstein, Rules and Institutions, Routledge.
Camerer, C.F., 2011. Behavioral Game Theory: Experiments in Strategic Interaction,
Princeton University Press.
Cubitt, R.P. & Sugden, R., 2003. Common Knowledge, Salience and Convention: A
Reconstruction of David Lewis’ Game Theory. Economics and Philosophy, 19(02),
p.175‑210.
Elster, J., 1982. The Case for Methodological Individualism. Theory and Society, 11(4),
p.453‑482.
Geanakoplos, J., Pearce, D. & Stacchetti, E., 1989. Psychological games and sequential
rationality. Games and Economic Behavior, 1(1), p.60‑79.
Gintis, H., 2009. The bounds of reason: game theory and the unification of the behavioral
sciences, Princeton University Press.
Gintis, H., 2010. Social norms as choreography. Politics, Philosophy & Economics, 9(3),
p.251‑264.
23
Greif, A., 2006. Institutions and the path to the modern economy: lessons from medieval
trade, Cambridge University Press.
Hargreaves-Heap, S. & Varoufakis, Y., 2004. Game Theory: A Critical Introduction 2e éd.,
Routledge.
Hédoin, C., 2012. Linking institutions to economic performance: the role of macro-structures
in micro-explanations. Journal of Institutional Economics, 8(03), p.327‑349.
Hédoin, C., 2013a. Collective Intentionality in Economics: Making Searle's Theory of
Institutional Facts Relevant for Game Theory. Erasmus Journal for Philosophy and
Economics, forthcoming.
Hédoin, C., 2013b. A Framework for Community-Based Salience: Common Konwledge,
Common Understanding and Community Membership. Working paper. University of
Reims Champagne-Ardenne, France.
Lewis, D.K., 1969. Convention: a philosophical study, John Wiley and Sons.
Marmor, A., 2009. Social Conventions: From Language to Law, Princeton University Press.
Menzies, P. & List, C., 2010. The causal autonomy of special sciences. In C. Mcdonald & G.
Mcdonald (Eds.), Emergence in Mind, Oxford University Press.
Morris, S., 1995. The Common Prior Assumption in Economic Theory. Economics and
Philosophy, 11(02), p.227‑253.
Rabin, M., 1993. Incorporating Fairness into Game Theory and Economics. The American
Economic Review, 83(5), p.1281‑1302.
Rawls, J., 1955. Two Concepts of Rules. The Philosophical Review, 64(1), p.3.
Schotter, A., 1981. The Economic Theory of Social Institutions, Cambridge University Press.
Searle, J.R., 1995. The construction of social reality, Simon and Schuster.
Searle, J.R., 2010. Making the social world: the structure of human civilization, Oxford
University Press.
Smit, J.P., Buekens, F. & Du Plessis, S., 2011. What Is Money? an Alternative to Searle’s
Institutional Facts. Economics and Philosophy, 27(01), p.1‑22.
Sugden, R., 1986. The Economics of Rights, Cooperation and Welfare 2e éd., Palgrave
Macmillan.
Sugden, R., 2000. The motivating power of expectations. In J. Nida-Rümelin & W. Spohn
(Eds.), Rationality, Rules, and Structure, Kluwer Academic Publishers, p.103-125.
Taylor, M., 1987. The Possibility of Cooperation, Cambridge University Press.
Udehn, L., 2001. Methodological individualism: background, history and meaning,
Routledge.
24
Ullmann-Margalit, E., 1977. The emergence of norms, Clarendon Press.
Vanderschraaf, P., 1998. Knowledge, Equilibrium and Convention. Erkenntnis, 49(3),
p.337‑369.
Young, H.P., 1998. Individual strategy and social structure: an evolutionary theory of
institutions, Princeton University Press.
25